--- name: claude-md-reflect description: Analyze chat history to identify areas for improving CLAUDE.md instruction files. Use when the user says "reflect on CLAUDE.md", "improve CLAUDE.md", or requests analysis of instruction files after a work session. Reviews recent conversation for misunderstandings, inconsistencies, or areas where instructions could be clearer. --- # CLAUDE.md Reflection and Improvement ## Overview This skill analyzes recent chat history and existing CLAUDE.md files to identify improvement opportunities. It follows a three-phase workflow: Analysis, Interaction, and Implementation. ## When to Use Trigger this skill when the user requests: - "reflect on CLAUDE.md" - "improve CLAUDE.md" - Analysis of instruction files after encountering issues ## Workflow ### Phase 1: Analysis Use the Task tool with subagent_type='general-purpose' to analyze the entire session history. **Find all CLAUDE.md files:** ```bash python scripts/find_claude_md.py ``` The script returns paths to all CLAUDE.md files in the repository, sorted by depth (root first). **Analyze for:** - Inconsistencies in Claude's responses vs user expectations - Misunderstandings of user requests - Areas where Claude lacked necessary context - Repetitive corrections from the user - Common mistakes or confusion patterns - Instructions that belong in a different CLAUDE.md file **Read reference materials:** - Read `references/anthropic-best-practices.md` for official CLAUDE.md guidelines - Read `references/anti-patterns.md` for common mistakes to avoid ### Phase 2: Interaction Present findings to the user using the AskUserQuestion tool with checkboxes. **For each suggested improvement:** 1. Explain the issue identified in chat history 2. Propose a specific change (add, update, or delete content) 3. Show which CLAUDE.md file to update 4. Explain the rationale and expected impact **Format suggestions as:** - Clear, actionable changes - Specific text additions or modifications - NOT vague improvements **Example:** ``` Issue: Claude repeatedly used 'var' instead of 'const' in JavaScript code Proposal: Add to CLAUDE.md: "NEVER use var in JavaScript, use const or let" File: ./frontend/CLAUDE.md Rationale: Prevents repeated corrections for JavaScript variable declarations ``` Wait for user approval on each suggestion before proceeding to implementation. ### Phase 3: Implementation For each approved change: 1. **Identify correct file and section:** - Root CLAUDE.md: Project-wide guidance - Subdirectory CLAUDE.md: Component-specific guidance - User CLAUDE.md (~/.claude/CLAUDE.md): Personal preferences 2. **Update the file:** - Add, update, or delete content as approved - Keep changes minimal and targeted - Maintain existing structure and formatting - Use bullet points, not paragraphs - Add emphasis (NEVER, ALWAYS, IMPORTANT) where critical 3. **Follow anti-patterns guidance:** - NO hyperbolic language (crucial, proper, robust, excellent) - NO vague instructions (write good code, use best practices) - NO historical comments (we used to do X, now we do Y) - NO redundant information - KEEP it concise and actionable 4. **Verify hierarchy:** - Root CLAUDE.md should link to subdirectory CLAUDE.md files - Instructions should be in the most appropriate file for their scope ## File Organization **Root CLAUDE.md:** - Project-wide guidelines - Links to subdirectory CLAUDE.md files - Common commands and workflows **Subdirectory CLAUDE.md:** - Component-specific guidance - Should be referenced from root CLAUDE.md **Example root CLAUDE.md structure:** ``` # Project Instructions ## Subdirectories - Frontend: @frontend/CLAUDE.md - Backend: @backend/CLAUDE.md ## Commands - npm run build: Build all components - npm test: Run all tests ## Code Style - Use 2-space indentation - NEVER commit directly to main branch ``` ## Resources - `scripts/find_claude_md.py` - Locate all CLAUDE.md files in repository - `references/anthropic-best-practices.md` - Official Anthropic CLAUDE.md guidelines - `references/anti-patterns.md` - Common mistakes to avoid